Source-linked AI summary
Strategies for Structuring Story Generation
Angela Fan, Mike Lewis, Yann Dauphin
TL;DR
Long-story language models struggle with plot coherence and consistent entities when generating word by word. This paper decomposes generation into action planning, narrative realization, and context-sensitive entity reference generation, and human evaluations favor the resulting stories while analyses report greater event and entity diversity.
Problem
Language models struggle to coalesce phrases into coherent plots and maintain character consistency because they generate long stories at the word level.
Method
The model generates predicate-argument action sequences, realizes a placeholder-based story, and replaces entity placeholders with context-sensitive references.
Results
Human evaluators prefer the full decomposition over strong baselines, while action decomposition increases verb diversity and entity methods generate more unique entities.
Takeaways & Limitations
Separating actions and entities improves story quality according to human evaluation and automated metrics.
Abstract
from arXiv · showhide
Writers generally rely on plans or sketches to write long stories, but most current language models generate word by word from left to right. We explore coarse-to-fine models for creating narrative texts of several hundred words, and introduce new models which decompose stories by abstracting over actions and entities. The model first generates the predicate-argument structure of the text, where different mentions of the same entity are marked with placeholder tokens. It then generates a surface realization of the predicate-argument structure, and finally replaces the entity placeholders with context-sensitive names and references. Human judges prefer the stories from our models to a wide range of previous approaches to hierarchical text generation. Extensive analysis shows that our methods can help improve the diversity and coherence of events and entities in generated stories.
1 Introduction
The paper addresses long-story generation's difficulty with plot-level coherence and character consistency by decomposing generation into coarse-to-fine stages. Its action- and entity-focused approach improves event and entity diversity while receiving favorable human evaluations.
- Language models generate stories word by word, making high-level interactions between plot points difficult to capture.
- Coarse-to-fine decompositions first expose abstract representations, support stage-specific modeling, and require no manual labels.
- The paper evaluates keyword extraction, sentence compression, and summarization as representative hierarchical decompositions because good decomposition properties remain unclear.
- The proposed three-step method models action sequences, generates a placeholder-based narrative, and rewrites placeholders into context-sensitive entity references.
- Trained on 300k WRITINGPROMPTS stories, the approach improves generation, increases verb diversity, and produces context-appropriate novel names.
2 Model Overview
The model formalizes story generation as coarse-to-fine generation from an abstract representation, then trains the abstract and realization models separately. This makes the decomposed objective tractable while retaining a convolutional sequence-to-sequence architecture for long-range context.
- A story x is converted into a more abstract representation z, separating abstract planning from surface generation.
- Generation samples first from p(z), then from p(x|z), but marginalizing over z is generally intractable.
- A deterministic posterior q(z|x) assigns each story a single z∗ obtained from semantic role labeling or coreference resolution.
- The resulting models p(z∗) and p(x|z∗) can be trained tractably and separately.
- The architecture uses convolutional encoder and decoder networks, encoder-decoder attention, and gated multi-head self-attention for previously generated words and long-range context.
3 Modelling Action Sequences
Action-sequence modeling uses semantic role labeling to represent stories through predicates and arguments, emphasizing logical event structure. A dedicated verb-attention head further conditions generation on earlier verbs and reduces repetitive, generic event choices.
- Semantic role labeling identifies predicates, arguments, and semantic roles, abstracting over different expressions of the same semantic content.
- The model generates an SRL decomposition from the prompt by concatenating predicates and arguments, separating sentences with delimiters, and retaining core arguments.
- Verb Attention Mechanism: A specialized verb-attention head attends only to previously generated verbs, using a zero vector before any verb exists.
- Verb Attention Mechanism: Verb attention generates a more diverse array of verbs and reduces repetition in generation.
4 Modelling Entities
Entity modeling addresses rare names and inconsistent references by anonymizing mentions before generation and restoring context-sensitive surface forms afterward. The system combines clustering, sequence-to-sequence reference generation, pointer copying, and subword modeling.
- Entity names are rare tokens, and one entity may appear through multiple strings, making character consistency difficult for language models.
- The method replaces coreferent mentions with placeholders, then uses a subword sequence-to-sequence model to restore references from context.
- The paper compares NER-based anonymization, which clusters identical strings, with coreference-based anonymization, which clusters different mentions of the same entity.
- Reference generation uses placeholder and story context, with pointer copying to encourage consistent reuse of entity names.
- Sub-word Modelling: Subword modeling is evaluated against word-based and character-level entity-generation models because entities are often rare or novel words.
- Coreference-based Entity Reference Generation: Coreference-based generation additionally uses local bag-of-words context and previously generated references to choose names, pronouns, or nominal expressions consistently.
5 Experimental Setup
The experiments evaluate story-generation decompositions against direct generation and representative hierarchical baselines using human judgments and automated measures.
- Dataset: The study uses WRITINGPROMPTS, a dataset of 300k stories paired with premises, with stories averaging 734 words.Stories are limited to 1,000 words, with fixed vocabularies for prompts and stories.
- Baselines: Baselines decompose generation through summarization, keyword extraction, and sentence compression, alongside the Fusion model’s direct prompt-to-story generation.The keyword baseline extracts n = 10 keyword phrases per story.
- Evaluation: Human evaluations compare different decomposed models for story generation, including models using SRL action plans and coreference-based entity clusters.The evaluation examines whether these decompositions produce stories preferred by human judges.
- Automated Evaluation: Automated evaluation measures the negative log likelihood of different decomposition stages, where lower values indicate easier modeling.Stage 1 generates an intermediate representation, while Stage 2 generates the story conditioned on it.
- Automated Evaluation: Generated-story similarity to training stories is assessed using maximum and average longest common subsequences of tokens.This measure is used to examine copying from human-written training stories.
- Implementation and Generation: Models are implemented in fairseq-py using a convolutional architecture, with stories sampled at temperature 0.8 and top-k value k = 10.Evaluations require stories to contain at least 150 words and truncate stories longer than 250 words at the nearest sentence.
6 Experiments
Experiments show that decomposing stories into structured actions and anonymized entities improves human-rated quality, event diversity, entity generation, and coherence relative to baselines, though gaps remain versus human writing.
- 6.1 Comparing Decomposition Strategies: The proposed decomposition copies slightly shorter sequences from the training set than baselines.This comparison uses maximum and average longest common subsequences of generated stories with human-written training stories.
- 6.1 Comparing Decomposition Strategies: Human evaluators preferred the full decomposition over baselines, including a carefully tuned Fusion model, in blind comparisons.The strongest model used SRL-based action plans and coreference-based entity anonymization.
- 6.2 Effect of SRL Decomposition: SRL predicate-argument decomposition improves verb diversity, while verb attention provides further improvement and reduces repetition.Models can often outline clear action sequences, but all remain below human story diversity.
- 6.3 Comparing Entity Reference Models: Character-level entity modelling outperforms BPE and word-based models because entity names are diverse.Entity-reference evaluation compares accuracy for first mentions and subsequent references.
- 6.3 Comparing Entity Reference Models: Full-story context gives the best entity-reference ranking accuracy, exceeding models with no context or only leftward context.Complete-story context is available because generation is multi-stage.
- 6.4 Effect of Entity Anonymization: The proposed decompositions generate more unique entities than strong baselines, although coreference anonymization produces fewer unique names than independent naming.Coreference models reuse previous names more often and use more pronouns.
- 6.4 Effect of Entity Anonymization: The full model produces more non-singleton coreference chains and more diverse mentions within clusters, but both measures remain below human-story levels.These analyses assess consistency and diversity of entity references.
7 Related Work
The paper situates its approach among classical multi-stage text generation, story-planning systems, entity language models, and non-autoregressive generation, emphasizing planning rather than faster decoding.
- Related Work: Classical text-generation systems separated content and discourse planning, sentence aggregation and lexicalization, and referring-expression resolution.The proposed work builds upon these staged approaches.
- Story Generation with Planning: Prior story-generation systems used handcrafted character-action models, story events, plot graphs, summaries, fragments, and image sequences.These approaches represent story planning at several levels of abstraction.
- Story Generation with Planning: Recent decomposition methods extract skeletons or storylines and condition generation on them, using reinforcement learning or manual storyline supervision in some cases.The paper evaluates alternative decomposition strategies to understand their properties.
- Entity Language Models: Entity language models have addressed entity types, representations, dynamic updates, and discourse-based reference tracking.These strands of work motivate explicit handling of entities in story generation.
- Non-Autoregressive Generation: Non-autoregressive models generate templates, insert words in arbitrary order, or decode in parallel, whereas this paper decomposes generation to support planning rather than inference efficiency.The distinction is between structural planning and faster decoding.
8 Conclusion
The paper concludes that separating action and entity generation provides an effective approach to short-story writing, improving story quality in human and automated evaluations.
- Conclusion: Separating action and entity generation improves short-story quality according to human evaluation and automated metrics.The conclusion summarizes the proposed decomposition as effective for writing short stories.